Double Mixture: Towards Continual Event Detection from Speech